Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Commits on main
Studio: count the LTX-2.5 encoder worker context in every whole-module phase (#15151)
d4b0044 Daniel Han committed 28 days ago
FP8 block GEMM: find torchao 0.18 blockwise_fp8_gemm and route each call to the faster kernel (#15153)
8f8eaee Daniel Han committed 28 days ago
Studio video: decode the Wan VAE untiled or in the fewest tiles that fit, also when offloaded (#15148)
ae0b163 Daniel Han committed 28 days ago
Studio diffusion: read the local int8 / fp8 prequant cache back on pipeline loads (#15146)
25a162d Daniel Han committed 28 days ago
Honour use_gradient_checkpointing=False over TRL's gradient_checkpointing=True default (#15145)
40840d7 Daniel Han committed 28 days ago
Studio diffusion: keep SDXL resident on 15-16 GB cards from its measured activations (#15144)
e3d92b4 Daniel Han committed 28 days ago
Studio: show prompt progress, live generation details and where a GGUF loaded (#14909)
e139d53 oobabooga committed 28 days ago
Studio: centralize GGUF identity rules and exclude incomplete GGUF splits (#15073)
0edb8f5 oobabooga committed 28 days ago
FSDP2: replicated grad sync without host syncs, once per gradient accumulation cycle (#15135)
065955d Daniel Han committed 28 days ago
Cross entropy: 8192-wide chunks at 4 warps for the chunked forward on B200 (#15141)
6eb93c8 Daniel Han committed 28 days ago
Remove hidden host-device syncs from per-step training paths (#15133)
50bb320 Daniel Han committed 28 days ago
FP8 block GEMM: TMA descriptor kernel with a 4-stage K loop on B200 (#15136)
8f6e39d Daniel Han committed 28 days ago
Studio diffusion: keep whole-module offload for a bf16 denoiser past the streamed floor (#15129)
609524d Daniel Han committed 28 days ago
Preserve runtime settings when unsloth start names the loaded model (#15026)
c3280b2 Nilay committed 28 days ago
Context parallel RL: single-tensor all-gather for hidden states and reduce-scatter in backward (#15132)
e2ba4af Daniel Han committed 28 days ago
Studio: remember chat preferences per account across logins (#15025)
1052e93 Nilay committed 28 days ago
Preserve padding masks in distillation backbone forwards (#15068)
98c1fdc cherry77-cloud committed 28 days ago
Training SDPA: bucket sequence lengths where torch picks cuDNN, so padded batches stop rebuilding plans (#15134)
a224ff3 Daniel Han committed 28 days ago
Explain disabled SentencePiece before loading MOSS-TTS-Nano (#14693)
4902adf Lee Jackson committed 28 days ago
Add the Unsloth provider to OpenClaw’s selected configuration (#15028)
9a72a04 Nilay committed 28 days ago
Sync before re-reading a buffer stored in an Inductor reduction loop (torch < 2.14) (#15000)
082c52a Daniel Han committed 28 days ago
Studio: Fix model pickers opening an empty On Device tab (#14985)
cc448f9 Nilay committed 28 days ago
Keep managed vLLM and SGLang environments usable across Docker updates (#15024)
263af9c Nilay committed 28 days ago
Run Unsloth's patched forwards only on Unsloth-loaded models (#15101)
8c69315 Daniel Han committed 28 days ago
Restore the hub source and endpoint env between backend tests (#15131)
e58dd25 Daniel Han committed 28 days ago
Studio video: pin MiniMax-H3's lazily streamed blocks in place so the stream prefetch engages (#15118)
d32a267 Daniel Han committed 28 days ago
MXFP4 torch decode: fp16 in fp32 on torch < 2.12, where ldexp overflows (#15121)
edf19f5 Daniel Han committed 28 days ago
Preserve reasoning limits when unsloth start changes context (#15027)
4c4ef58 Nilay committed 28 days ago